In progress — ICAETM 2026 submission
Time-Series Modeling of Pulsed-Mode N₂O RCS Thrusters
A lumped-parameter physics model of a pulsed nitrous-oxide monopropellant thruster, turned into a time series, characterized statistically, and forecast with an LSTM.
Inside the work
Every pulse leaves a memory.
Illustrative N₂O feed, catalytic decomposition, nozzle expansion, and pulse-to-pulse thermal memory; the timing and amplitude are schematic.
Open the valve
A brief propellant feed starts a new pulse through the catalyst bed.
Decompose and expand
Catalytic decomposition raises chamber pressure; the products expand through the nozzle to generate thrust.
Carry the heat
The bed retains heat between pulses, linking the next response to the previous one.
Overview
Pulsed-mode N₂O monopropellant reaction-control thrusters fire in short duty-cycle bursts, and each pulse leaves thermal state (catalyst-bed temperature, chamber soak-back) behind for the next one. This project builds a lumped-parameter physics model of catalytic N₂O decomposition, drives it through a realistic pulse-train duty cycle, and treats the resulting per-pulse impulse-bit / chamber / bed measurements as a time series rather than a set of independent samples.
The pipeline is five layers deep: a physics model of a single pulse, a pulse-train generator that carries thermal state forward between pulses (soak-back), a statistical characterization pass (drift, periodicity, stationarity), predictive models benchmarked against a classical baseline, and a six-figure, paper-quality visualization suite.
Dataset
600 pulses
Lognormal duty-cycle dataset
400 pulses
Uniform duty-cycle dataset
16 figures
PNG @300 DPI + SVG, regenerable
Chart
LSTM's RMSE improvement over the ARIMA(1,0,1) baseline
Pipeline
0 · Literature grounding
Real parameter ranges for N₂O monopropellant thrusters, encoded as defaults in config.py.
1 · Physics model
Lumped-parameter single-pulse model of chamber pressure, chamber temperature, and catalyst-bed temperature.
2 · Pulse-train time series
A duty-cycle engine fires a pulse train and carries thermal state forward between pulses (soak-back), producing 600 lognormal-spaced and 400 uniform-spaced pulse datasets.
3 · Time-series characterization
Drift, Mann-Kendall trend test, Augmented Dickey-Fuller stationarity, autocorrelation, and FFT spectral analysis on every feature.
4 · Predictive models
An LSTM forecaster evaluated out-of-sample against a classical ARIMA(1,0,1) baseline on RMSE/MAE.
5 · Visualization suite
16 paper-quality figures (PNG @300 DPI + matching SVG) plus a graphical abstract and animations.
Results
Time-series characterization — 600-pulse dataset
| Feature | Trend (Mann-Kendall) | Kendall τ | Trend p | Stationary (level) |
|---|---|---|---|---|
| Mean thrust | Increasing | 0.20 | 0.0 | True |
| Peak chamber pressure | Increasing | 0.195 | 0.0 | True |
| Bed start temperature | Increasing | 0.51 | 0.0 | True |
| Minimum impulse bit (MIB) | No trend | 0.019 | 0.494 | True |
Comparison
LSTM vs. ARIMA(1,0,1) out-of-sample forecast error
| Target | ARIMA RMSE | LSTM RMSE | RMSE improvement |
|---|---|---|---|
| Bed start temperature | 2.0298 | 0.7892 | 61.1% |
| Peak chamber pressure | 15214.3 | 14463.1 | 4.9% |
| Mean thrust | 0.2254 | 0.2078 | 7.8% |
Key Findings
What the work shows
Bed start temperature is where the LSTM earns its keep — 61.1% RMSE improvement over ARIMA, because soak-back drift between pulses is a nonlinear, state-dependent effect a linear ARIMA model can't capture.
Mean thrust and peak chamber pressure are only mildly improved by the LSTM (4.9–7.8%) — these are closer to stationary, weakly-drifting series where ARIMA is already competitive.
All four characterized features are stationary in levels (ADF), and three of four show a statistically significant increasing drift (Mann-Kendall p = 0.0) — consistent with progressive catalyst-bed thermal loading across a 600-pulse duty cycle.
Every dataset, table, and figure in the repo is regenerable from the analysis/ CLI suite with a fixed random seed.